Atrial fibrillation: an AI finds its trace on an ECG that looks normal

Why does atrial fibrillation so often go unnoticed?

Atrial fibrillation is a rhythm problem: the atria, the upper chambers of the heart, beat in a chaotic way. It raises the risk of stroke, heart failure and death.

It often causes no symptoms, and it can come and go. An electrocardiogram (ECG), the test that records the heart's electrical activity, taken during a calm spell will look normal. Finding it often means wearing a recorder for days, which is costly and picks up few cases.

What did the researchers do?

The team at the Mayo Clinic in the United States trained a convolutional neural network, a type of AI that is very good at recognizing patterns in signals and images. They gave it standard 10-second, 12-lead ECGs, all with a normal rhythm, recorded between 1993 and 2017.

Patients with at least one ECG showing atrial fibrillation or atrial flutter (a related problem) were counted as positive. In total, 649,931 ECGs from 180,922 patients were split into three groups: one to learn, one to tune the tool and one to test it.

The question was simple: does an ECG that looks normal keep a trace of fibrillation that the eye cannot see?

What were the results?

In the test group of 36,280 patients, a single ECG was enough for the AI to identify atrial fibrillation with an area under the curve of 0.87, on a scale where 1 would be perfect. Its sensitivity, the share of patients with the condition it picks up, was 79.0%, and its overall accuracy 79.4%.

When all the ECGs taken during the first month of each patient's study window were used, the area under the curve rose to 0.90 and accuracy to 83.3%.

The authors conclude that an ECG taken during a normal rhythm can identify, at the point of care, people who have atrial fibrillation.

What could this change for patients?

An ECG is quick, cheap and available almost everywhere. If this approach holds up, a routine test could flag the people who need a longer search for fibrillation, instead of giving everyone a recorder.

But finding the condition is not enough on its own: it is the treatment a doctor then decides on that can lower the risk of stroke.

What are the limits?

This is a retrospective analysis of past ECGs from a single large center. It needs to be confirmed in other hospitals and other populations.

The tool also makes mistakes: roughly one ECG in five is put in the wrong group, one way or the other. So an alert is not a diagnosis, and it still has to be shown that using the tool prevents strokes.